About Cloud TPU reservations

This document explains how to reserve Cloud TPUs for your exclusive use by requesting one of the following:

Reservations are useful for the following cases:

  • Planned or unplanned usage spikes
  • Obtaining high-demand resources
  • Long-running training jobs and inference workloads
  • Workloads requiring a high assurance of capacity

A reservation is one consumption option for Cloud TPU. For more information, see Cloud TPU consumption options.

Choose a reservation type

To choose the best reservation type for your needs, consider the duration of your workload and whether you want to commit to a longer-term usage agreement in exchange for higher discounts:

  • Short-term workloads: Choose a future reservation for up to 90 days (calendar mode) if you need capacity for planned, time-bound events, such as short-term training tasks or seasonal spikes in user demand.
  • Long-term workloads: Choose a future reservation for one year or longer if you have steady, predictable workloads that run continuously and want to take advantage of committed use discounts (CUDs).

Both reservation types give you a high level of assurance that TPUs are available when you need them, for a specified time period. The following table shows the differences between the reservation types:

Future reservations for up to 90 days (in calendar mode) Future reservations for one year or longer
Duration 1 to 90 days 1 to 3 years
Supported TPU versions v5e, v5p, v6e, and TPU7x All TPU versions
Cost (for more information, see DWS pricing) Up to 30% less than on-demand 30% to 55% less than on-demand
How to request Self-service using the Compute Engine API or the Google Cloud console Manual process through Cloud Sales or your account manager
Committed use discount (CUD) Not supported CUD required

Create a reservation

Depending on the reservation type you choose, you request and create the reservation in one of the following ways:

What's next